Work
Research and applied builds, in various states of finished.
An 8-step agentic LLM pipeline that helps instructors plan classroom role-play activities, with a multi-layer review system to catch confident-but-wrong AI output and every answer grounded in the real McGraw Hill textbook via Postgres/pgvector retrieval.
Five progressively optimized CUDA kernels — culminating in a Kogge-Stone parallel-prefix-scan with warp-shuffle intrinsics — that parallelize Mamba's inherently sequential selective-scan recurrence, cutting inference from 246ms to 1.6ms per forward pass (156x) on an A100.
A real-time forecasting system for Tallinn's Old Town that streams pedestrian-sensor and weather data via Kafka, forecasts traffic 8 hours ahead with MSTL/AutoARIMA, and re-architects onto serverless AWS (Lambda + EventBridge ingestion, RDS Postgres, Fargate) for on-demand retraining and dashboard hosting.
A multilingual speech-to-speech translation app built at HackArizona — chains OpenAI Whisper (ASR), Meta NLLB-200, and MMS-TTS into a low-latency, chunk-based pipeline supporting 12+ languages, shipped as a live hold-to-speak demo over Flask + Socket.IO.
A Python/tkinter desktop app that logs study session start/stop times directly to a Notion "Time" database via the Notion API, replacing manual tracking. Built, debugged, and confirmed working on Windows.
Built for the e-Yantra Robotics Competition (IIT Bombay) — Verilog HDL control logic on a Cyclone IV FPGA (FSMs, PWM motor drivers, UART, ADC) plus embedded AVR C firmware fusing IR and color sensors for closed-loop, obstacle-aware path tracking.